The Next Step in E-diagnostics: Mining the Tool Sensors

Effective process learning and process control define the cost-competitiveness of advanced fabs today. Delivery interruptions are common because of process and yield variations within a fab. It often takes too long to diagnose the root cause of such variations, which persist as more experiments are run in an attempt to isolate the variable output and throughput conditions. As more and more engineering resources are engaged in "fire drills" to resolve production variations, they are not available to determine process windows and perform continuous improvement. Accordingly, the timely qualification of process technologies for-next generation devices is at risk if process learning cannot be a core skill set within a fab.

Traditional e-diagnostics systems monitor tool performance and provide "maintenance needed" alerts to service and engineering personnel, most often after a failure has already occurred. Much more can be derived from the rich sensor data generated as wafers pass through process chambers. This "next step" in e-diagnostics is to employ proven enterprise data mining (EDM) techniques to correlate device yield and performance with the vast amount of tool-level and wafer-level chamber sensor data (Fig. 1 ).

With this new approach, yield and process-level issues can be uncovered down to a particular sensor reading on a specific tool process chamber. Once a specific tool issue can be identified to have an impact on process results, specific e-diagnostics monitors can be targeted to prevent future yield and process excursions, completing a closed-loop process learning effort.

1. Block diagram of Applied Materials’ Sensor Data Mining services.

In step one, a massive amount of diverse defects, metrology, e-test, probe and yield data, plus process and MES flow historical information, must be gathered and fed into a set of common and linked databases to support critical metric monitoring. Hundreds to thousands of attributes, along with the historical process tool sensor data, need to be considered. Immense amounts of data are collected (Fig. 2 ); tool sensors alone can generate terabytes of data. This aggregation of data must be gathered from fabwide data sources and exported automatically into the correlation engine while maintaining the integrity of the information. Upon export to analysis tools, the data are parsed, filtered, sorted and prepared for an analysis engine.

2. Database size for a typical device.

Step two is to analyze the vast amount of data accumulated. The analysis uses proprietary data-mining algorithms to uncover which metrology metrics best correlate to factory metrics, such as yield or tool uptime and utilization. The algorithms have a drill-down capability to perform automated searching until the more actionable correlations are made. Initial analyses may link process performance to electrical parameters or in-line metrology data (defect levels, critical dimensions, etc.). The end result will be to identify the processes and tool attributes that have the largest effect on process performance. This automated analysis allows for potentially harmful tool and chamber mismatch conditions to be uncovered and ranked in order of impact. E-diagnostics with automated data mining processes can provide correlation results in less than 24 hours, compared with conventional manual trial-and-error design-of-experiments (DOE), which may take weeks.

Step three analyzes the sensor data to identify a specific root-cause issue. Once a specific tool and sensor has been identified as a potential culprit, visualization and analysis software allows for chamber-to-chamber and tool-to-tool comparisons and contrasts. Using these tools, the process and equipment engineers can quickly and precisely determine actionable data-driven optimization steps.

In step four, corrective and improvement action and subsequent validation completes the analysis-driven optimization process.

In step five, specific tool sensor limits at the recipe level can be configured via the e-diagnostics system as automatic monitors (upper and lower control limits) to prevent future yield and process excursions. This important step completes the process and excursion control loop (Fig. 3 ).

3. Sensor Data Mining methodology.

The application and results of the Sensor Data Mining methodology can best be illustrated by the following two plasma etch case studies. These studies highlight the e-diagnostics and Sensor Data Mining sequence: continuous data collection and storage, determination and prioritization of issues, analysis to determine the specific corrective actions, the verification of improvements from the appropriate implementation and, finally, set up of proper control limits to prevent future yield excursions.

Case study 1: M2-M3 contact resistance. An EDM yield analysis indicated that the metal 2 to metal 3 contact resistance was found to be a key yield detractor. Further drill-down analysis pointed to a specific via etch tool as the culprit. Stored e-diagnostics data was extracted for the three operating chambers on this tool (A, B and D) that were used during the time period of interest.

Automated sensor data analysis of the tool parameter data was conducted to determine if there was any correlation to the electrical test data. Two correlations emerged: low foreline pressure on the oxide via etch tool with high contact resistance (Fig. 4), and low foreline pressure with chambers A and D operation (Fig. 5 ).

4. Sensor Data Mining of the etch tool e-diagnostics sensor data identified a correlation between foreline pressure in the oxide etch tool and M2-M3 contact resistance. A low foreline pressure correlated with high contact resistance and poor yield.

5. Sensor Data Mining of the etch tool e-diagnostics data identified two mismatched process chambers on the oxide etch tool that had low foreline pressure.

Further data mining analysis pointed to chamber backpressure during a specific process step of the via etch process. This was traced to etch byproduct buildup on the throttle valve sealing surface. Corrective action was taken and the two chambers were requalified for production. Specific e-diagnostic control limits were set to monitor the chamber backpressure during the specific recipe step to prevent future yield excursions.

Case study 2: Polysilicon CD. In this case, e-diagnostics with automated Sensor Data Mining revealed a correlation between the final polysilicon line critical dimension (CD) and the process chamber pressure during one of the polysilicon etch steps (Fig. 6 ).

6. Automated sensor data mining revealed a correlation between chamber pressure during one of the polysilicon etch steps and the final polysilicon CD. The target CD is 0.12-0.13 µm with a chamber pressure reading of 150-250 mTorr.

Further data mining analysis revealed mismatched process chambers with regard to pressure at the critical etch step (Fig. 7 ). More effective wet cleaning reduced the mismatch, and the chamber pressure was retargeted to obtain the target CD.

7. Sensor Data Mining of the etch tool e-diagnostics data identified two mismatched process chambers on the oxide etch tool that had low foreline pressure.

Specific chamber pressure reading monitors upper and lower limits of 150 and 250 on this specific recipe-step were configured into e-diagnostics systems for closed-loop process control and yield loss prevention.

Sensor Data Mining techniques with tool-level e-diagnostics data can significantly reduce the cycle time for process, tool and chamber matching, which is critical to maintain process performance in today's aggressive and ever-shrinking process windows. As shown in the above examples, unmatched chamber conditions causing undesirable process variation can be quickly and precisely targeted, allowing for a rapid implementation of corrective action. The reduction in number of experiments needed and time-to-results is dramatic.

Typically, the turn-around-time takes many weeks, involving hours of process engineering labor, many test wafers, and hours of otherwise productive tool time to match multiple parameters, many of which may have no yield impact. In our cases, these tasks were reduced to less than one week. The tools and chambers were back in production quickly, enabling engineering resources to be dedicated to continuous improvement and new product introductions. Furthermore, the specific set of yield-sensitive tool sensors can now be continuously monitored via the e-diagnostics system to prevent related yield excursions from reoccurring. This technology can provide closed-loop, end-to-end yield and process excursion control, which can help to continuously eliminate yield excursions caused by unexpected tool variations, and reduce delivery interruptions to the fab's customers.

Mature e-diagnostics capabilities have been combined with proven enterprise data mining technology to pinpoint the specific critical process conditions and variables that affect process control. Sensor Data Mining combines fab-level data (particles, defects, metrology, e-test, yield and others) with e-diagnostics, and tool-based sensor data to identify significant correlations among all variable combinations. This combination of technologies can provide a comprehensive prioritized listing of the potential issues related to key tool parameters, including hard-to-find second-order effects.